Communications Laboratory,

  • 格式:pdf
  • 大小:139.55 KB
  • 文档页数:8

ScalabilityAnalysisofAudio-VisualPersonIdentityVerification

JacekCzyz1,SamyBengio2,ChristineMarcel2,andLucVandendorpe11CommunicationsLaboratory,

Universit´ecatholiquedeLouvain,B-1348Belgium,czyz@tele.ucl.ac.be,2IDIAP,CH-1920Martigny,

Switzerland{Samy.Bengio,Christine.Marcel}@idiap.ch

Abstract.Inthiswork,wepresentamultimodalidentityverificationsystembasedonthefusionofthefaceimageandthetextindependentspeechdataofaperson.Thesystemconciliatesthemonomodalfaceandspeakerverificationalgorithmsbyfusingtheirrespectivescores.Inordertoassesstheauthenticationsystematdifferentscales,theperformanceisevaluatedatvarioussizesofthefaceandspeechusertemplate.Theusertemplatesizeisakeyparameterwhenthestoragespaceislimitedlikeinasmartcard.Ourexperimentalresultsshowthatthemultimodalfusionallowstoreducesignificantlytheusertemplatesizewhilekeepingasatisfactorylevelofperformance.ExperimentsareperformedonthenewlyrecordedmultimodaldatabaseBANCA.

1IntroductionWiththeadventofdigitalcommunicationandinformationsociety,reliableanduser-friendlypersonalidentityverificationbecomesmoreandmoreindispensableandcritical.Biometrics,whichmeasuresaphysiologicalorbehaviouralcharac-teristicofaperson,suchasvoice,face,fingerprints,iris,etc,providesaneffectiveandinherentlymorereliablewaytocarryoutpersonalidentification[4].Sev-eralfactorsinfluencethechoiceofabiometrictraitforaparticularapplication.Amongthem,distinctivenessanduserfriendlinessarecertainlythemostimpor-tant.Fordistinctiveness,thebiometrictraitshouldbedistributedwithalargevarianceinsidethetargetpopulation.Atthesametime,itshouldideallyremainconstantforagivenperson,orvarywithasmallvariance.Asforuserfriendli-ness,thesensorsthatcapturethebiometrictraitsshouldinterferewiththeuseraslittleaspossible.Alsothetraitrecordingsshouldbedoneinanunconstrainedandcontactlessmanner.Thesetworequirementsareunavoidablycontradictory.Therefore,ithasbeensuggestedtocombineorfuseseveraleasilyacceptedbio-metrictraits,inordertoachieveanacceptablelevelofdistinctivenessanduserfriendlinessatthesametime.Thistechniqueisknownasmultimodalbiometrics.Theaptitudeofmultimodalbiometricsforincreasingcorrectverificationrates2overmonomodalbiometricshasbeendemonstratedinseveralpreviousstudies,(seeforexample[3],[5],[8]).Apromisingapplicationconsistsincombiningbiometricefficiencyandsmartcard(SC)security,bystoringtheusertemplateonaSC[9].SC’sallowthetem-platetobesecurelyprotectedandavoidstoringbiometricdatainacentralserver(centralstorageisnotwellacceptedbyusers).HoweverstoragespaceofSC’sandtransmissionspeedbetweenserverandSC’slimittheusertemplatesize.Itisthereforeimportanttoevaluateperformanceasafunctionofthetemplatesize.Inthiswork,wepresentanidentityverificationsystembasedonthefu-sionofthefaceimageandtextindependentspeechdataofauser.Weanalyseitsscalabilitybyevaluatingtheperformanceatdifferentusertemplatesizes.Althoughlessaccurate,textindependentspeakerverificationallowsmorevari-abilityinthecontentutteredbytheuser.Forthisreasonitismoreuser-friendlythantextdependentspeakerverification.Thesystempresentedismodular:eachmonomodalalgorithm(faceandspeech)outputsamatchingscorereflectingitsconfidenceinthepresumedidentity.Thematchingscoresarethenconciliatedusingafusionalgorithmwhichoutputsthefinalauthenticationdecision.Ouranalysisoftheexperimentalresultsonarealisticdatabaseshowsthatthefusedsystemface-speechrequiresamuchsmallernumberofparameterstorepresentauserthanthebestmonomodalalgorithmatthesameperformancelevel.Fusioncanthereforehelpinreducingthestoragespacerequiredforclientdataandthusimprovethescalabilityoftheverificationsystem.Thepaperisorganisedasfollows.Themonomodalalgorithmsandthefusiontechniquesemployedarepresentedinsection2.Insection3,thescalabilityanalysisisdescribed.Thedatabaseandtheexperimentalprotocolarepresentedinsection4.Wediscusstheresultsinsection5,andwedrawconclusionsinthelastsection.

2FusionofFaceandSpeakerVerificationAlgorithmsWhentheidentityofauserhastobeverified,speechandfacearerecordedandcomparedtopreviouslycreatedusertemplate.Ascorereflectingthequalityofthematchingbetweenthetemplateandthedatatoverifyiscomputed.Thefusionofthetwoscoresresultingfromthespeechandfacealgorithmsleadstothefinaldecision.Hereafterwedescribebrieflythespeakerandfaceverificationalgorithmsandthefusiontechniques.FaceVerificationAlgorithmThefirststepinvolveslocalisationandregistrationofthefacepartintheinputimage.Inourimplementation,wehaveskippedthisstepbymanuallylocatingtheeyecoordinatesintheimage.Whileoftendoneintheliterature,itbiasesoptimisticallytheverificationperformance.Afterlocalisation,thefaceimageiscroppedandhistogramequalisationisappliedtoreducetheeffectoflight-ingvariation.TheFisherfaceapproach[1]isusedtoextractfeaturesfromthegraylevelfaceimage.ThisfeatureextractiontechniqueisbasedonPrincipalComponentAnalysis(PCA)andonLinearDiscriminantAnalysis(LDA).LDA